Concept Bottleneck Models

TL;DR

The paper proposes concept bottleneck models that constrain predictive reasoning to an intermediate, human-specified concept layer, enabling direct interpretation and test-time interventions. By training models to map x to c and then c to y, they demonstrate competitive task performance on knee osteoarthritis grading and bird identification while achieving high concept accuracy. Importantly, the authors show that intervening on predicted concepts at test time can significantly improve accuracy, and that the approach can enhance robustness to background shifts. The work also analyzes the trade-offs between training schemes (independent, sequential, joint) and emphasizes that concept supervision facilitates richer human-model collaboration in high-stakes domains like medicine. Overall, concept bottleneck models offer a practical path to interpretable, interactable AI without sacrificing accuracy, albeit at the cost of requiring concept annotations during training.

Abstract

We seek to learn models that we can interact with using high-level concepts: if the model did not think there was a bone spur in the x-ray, would it still predict severe arthritis? State-of-the-art models today do not typically support the manipulation of concepts like "the existence of bone spurs", as they are trained end-to-end to go directly from raw input (e.g., pixels) to output (e.g., arthritis severity). We revisit the classic idea of first predicting concepts that are provided at training time, and then using these concepts to predict the label. By construction, we can intervene on these concept bottleneck models by editing their predicted concept values and propagating these changes to the final prediction. On x-ray grading and bird identification, concept bottleneck models achieve competitive accuracy with standard end-to-end models, while enabling interpretation in terms of high-level clinical concepts ("bone spurs") or bird attributes ("wing color"). These models also allow for richer human-model interaction: accuracy improves significantly if we can correct model mistakes on concepts at test time.

Figures (5)

• Figure 1: We study concept bottleneck models that first predict an intermediate set of human-specified concepts $c$, then use $c$ to predict the final output $y$. We illustrate the two applications we consider: knee x-ray grading and bird identification.
• Figure 2: Left: The shaded regions show the optimal frontier between task vs. concept error. On OAI, we find little trade-off; models can do well on both task and concept prediction. On CUB, there is some trade-off, with standard models and joint models that prioritize task prediction (i.e., with sufficiently low $\lambda$) having lower task error. For standard models, we plot the concept error of the mean predictor (OAI) or random predictor (CUB). Mid: Histograms of how accurate individual concepts are, averaged over multiple random seeds. In our tasks, each individual concept can be accurately predicted by bottleneck models. Right: Data efficiency curves. Especially on OAI, bottleneck models can achieve the same task accuracy as standard models with many fewer training points.
• Figure 3: Successful examples of test-time intervention, where intervening on a single concept corrects the model prediction. Here, we show examples from independent bottleneck models. Right: For CUB, we intervene on concept groups instead of individual binary concepts. The sample birds on the right illustrate how the intervened concept distinguishes between the original and new predictions.
• Figure 4: Test-time intervention results. Left: Intervention substantially improves task accuracy, except for the control model, which is a joint model that heavily prioritizes label accuracy over concept accuracy. Mid: Replacing $c \to y$ with a linear model degrades effectiveness. Right: Intervention improves task accuracy except for the joint model. Connecting $c \to y$ to probabilities rescues intervention but degrades normal accuracy.
• Figure 5: In the TravelingBirds dataset, we change the image backgrounds associated with each class from train to test time (illustrated above for a single class).

Theorems & Definitions (6)

• Proposition 1: Relative excess error of independent bottleneck models vs. standard models in linear regression
• Lemma 1: Risk of the independent bottleneck model
• proof
• Lemma 2: Risk of the standard model
• proof
• proof : Proof of Proposition 1